EuSpRIG - Dr Simon Thorne
banner
eusprig.bsky.social
EuSpRIG - Dr Simon Thorne
@eusprig.bsky.social
110 followers 300 following 190 posts
Programme chair for The European Spreadsheet Risks Interest Group, EuSpRIG - (“yewsprig”) for short. Academic and researcher in software and spreadsheet quality
Posts Media Videos Starter Packs
7. MI5's bug the wrong people
MI5 collected the wrong phone data 1,061 times
Why? A formatting error in Excel requested numbers ending in “000”
#YouOnlyFormatTwice
6. Excel audit trail catches fraudster
A Harvard Prof. falsified data in behavioural studies.
It was Excel’s audit trail that exposed her.
Sometimes spreadsheets fight back.
#Integrity #Audit
5. The gene name tragedy
Scientists keep naming genes things like “MARCH1”
Excel helpfully turns them into dates.
20 years on, the problem still breaks research papers.
#Bioinformatics #ExcelFails
4. Courtroom chaos
During the Proud Boys trial, prosecutors accidentally shared classified messages.
They’d filtered them out — but they remained in hidden rows.
Discovered live in court.
#LegalFail #FOI
3. $92M on the line
Norway’s $1.5 trillion sovereign fund lost $92M.
All because someone entered the wrong date in a benchmark spreadsheet. One cell. Nine figures.
#FinanceRisk #Excel
2. Active MI5 staff compromised
Police in Northern Ireland accidentally published personal data of 10,799 staff, incl. MI5 operatives.
A hidden tab in Excel. A massive breach.
Cost: £200m+.
#DataRisk #SpreadsheetFail
Why you should care about spreadsheet risk – and come to #EuSpRIG2025

Every year, billions depend on spreadsheets. And every year, they go horribly wrong.

Here are 7 real cases that prove why #ExcelFails matter.👇
Out of interest, would a day workshop run by EuSpRIG on the safe use of #LLMs/#GenerativeAI for #spreadsheet modelling be of interest to people? We could do an in person or online workshop
Excited about my new paper: Large Language Models for #Spreadsheets introduces the #FLARE #benchmark, a deep dive into how #LLMs handle real spreadsheet logic, symbolic reasoning and error detection.
They sound confident, but still make dangerous mistakes. www.eventbrite.co.uk/e/eusprig-20...
EuSpRIG 2025 Annual Conference: Spreadsheet Productivity and Risks
Join us at EuSpRIG 2025 for insights on boosting your spreadsheet skills while managing risks - it's gonna be epic!
www.eventbrite.co.uk
"The Devil's guide to Vibe Driven Development" A chaotic, hilarious manifesto for anyone who's ever coded by gut feeling and vibes alone. You’ve done agile. Now try unhinged. Be careful with #LLMs and always validate and verify

recondite-basket-7d0.notion.site/The-Devil-s-...
The Devil's Guide to Vibe-Driven Development 👹 | Notion
1. Just Code by Vibes 🎯
recondite-basket-7d0.notion.site
decision makers using these spreadsheets second hand did not realise that the data may be limited or the assumptions too narrow for general purpose. All of this could have been avoided with a better system to analyse the results in the first place.
scientists were not able to run ad-hoc queries on the database containing results, so the data was exported to Excel for analysis (sounds familiar), what was not realised was that these spreadsheets were for very specific purposes with specific assumptions...
AI could be as transformative to audit as the spreadsheet was to accounting. But this will only happen if firms integrate it with care—balancing automation with accountability, and innovation with ethics.
EY, Accenture, Grant Thornton and others are investing heavily, partnering with major tech providers like Microsoft, Nvidia, and Google to embed AI across audit services. Falling behind could mean lower quality outcomes and reduced competitiveness.
Ethical deployment of AI requires:
– Transparent algorithms
– Bias audits and explainability
– Consistent data quality standards
– Clear escalation protocols
Without these, AI’s credibility in audit work is at risk
AI trained on past fraud may fail to detect new, creative fraud techniques. Over-reliance can dull professional skepticism. Leading firms stress the importance of human oversight and contextual understanding.
AI models can absorb historical biases—especially if the training data includes discriminatory decisions. For example, some LLM-based tools have shown racial bias in mortgage approvals. Auditing tools face similar risks of embedding flawed assumptions.
AI enables full-population testing, not just samples. It can detect subtle patterns, suggest strategic insights, and even scan past engagements to strengthen pitches and planning.
Thomson Reuters beta testers reported halved sample sizes and faster audit testing. Deloitte suggests AI could reduce audit costs by 25% through automation of routine work.
EY trialled an AI fraud-detection tool with 10 UK clients. It flagged issues in two cases—both later confirmed as fraud. That's a strong signal that AI can detect what human sampling might miss.
Recent audit failures have exposed the limits of traditional methods. Auditors need tools that can detect anomalies across entire datasets. AI is stepping into that role. Although my own testing suggests that this is still quite limited - check my forthcoming 2025 EuSpRIG paper